Skip to content
Other

SpanCall Lean v3

fast non-autoregressive tool calling

2 min readupdated

SpanCall Lean v3 (spancall_lean_modernbert architecture) by convoicon handles tool call detection and extraction. Instead of slowly generating JSON token by token like standard LLMs, it extracts the target function and its arguments in a single forward pass. Built on the bidirectional answerdotai/ModernBERT-large encoder with two specialized heads (H1 for tool selection and an H3 pointer head for parameter span extraction), it is ideal for AI agent and router developers who need low-latency user query parsing.

Key features

  • Single forward pass: constructs a ready-to-use function call without autoregressive JSON generation.
  • ModernBERT-large backbone: uses a bidirectional encoder supporting context lengths up to 4096 tokens (max_seq_len).
  • Two-step extraction: runs in two_step mode using a paired H1 tool head and H3 pointer head.
  • Single-call focus: configured to extract one tool call at a time (Single-call only: True) with a min_joint_margin: 0.02 threshold.

What you can do

  • Speed up voice and chat assistants: replace slow JSON generation when routing user commands to APIs.
  • Build long-context AI agents: process conversations and tool schemas up to 4096 tokens in a single pass.

Example

from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("convoicon/spancall-lean-v3")
cfg = json.load(open(hf_hub_download("convoicon/spancall-lean-v3", "resolve_config.json")))
weights = hf_hub_download('convoicon/spancall-lean-v3', 'model.pt')
# далее инициализируются SpanCallTransformerEncoder и LeanSpanCallModel
Read moreDetails
hermes_eval benchmark (limit=2000, n_F=1157): exact_F — 0.681, hit_F — 0.872, param_F1_F — 0.83, exact_all — 0.509.
For comparison, the B1a regex baseline scores exact_F 0.022 and param_F1_F 0.03 on the same data.
Limitations: supports only a single tool call per request (Single-call only), and running it requires recreating the model classes from the author's notebook.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.5 GB · ready files2.1 GB
  • FP8 / INT8recommendedweights 0.7 GB1.3 GB
  • 4-bitweights 0.4 GB1.0 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # convoicon/spancall-lean-v3
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download convoicon/spancall-lean-v3 --local-dir ~/models/spancall-lean-v3
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to spancall-lean-v3
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of spancall-lean-v3 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

convoicon/spancall-lean-v3
Author
convoicon
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments